2 p.m. - 3 p.m. Location: FN 2.102
Political Sciences Department, UT Dallas
Social Science Applications of Bayesian Structural VAR Models and Natural Language Processing
Bayesian Structural VAR Models and Natural Language Processing have been underutilized in Social Science research. This project studies the dynamic endogenous interactions of energy sector campaign contributions to members of the U.S. Senate Committee on Energy and Natural Resources and their reactions in the 105th-112th Congresses. Congressional rhetoric is segregated from the database of Committee hearings using a Java script, and then analyzed using a supervised Bayesian automated nonparametric natural language processing algorithm. The natural language processing of the speeches produces multiple time-series variables of the changes in the legislative rhetoric (organized into categories according to the exhibited attitude) on the level of each individual Senator. The densities of the categorized rhetoric over time are then linked to the campaign contributions data. The final stage of the analysis includes various Bayesian models - exogeneity tests and Markov-Switching for the hypotheses testing, as well as Bayesian vector autoregression for the analysis of the temporal dynamics within the congressional committee speeches. Conditioning on party and Senate class, I find strong evidence for a fully dynamic model, where rhetoric and contributions engage in a complex system of leads, lags, and contemporaneous interactions.
Sponsored by the Department of Mathematical Sciences